It is known that one of the tests performed in the preliminary steps for checking a permanent memory device is a test which allows to identify the cells of memory devices that are not erased sufficiently by an exposure to ultraviolet rays or that lose their erased state for various reasons.
This test consists in comparing the threshold in all the cells of the memory matrix with the threshold of an appropriate reference cell by means of a normal read operation performed with a sense circuit constituted by a sense amplifier which receives in input a signal from a cell being tested and a signal from the reference cell.
In this case, the reference cell has a threshold which identifies it as neither programmed nor virgin.
The cells compared with the reference cell must therefore have a lower threshold in order to assuredly be virgin cells.
The memory cells rejected as a consequence of this check are made redundant by means of the redundancy lines, which therefore allow to restore the "damaged" cells and therefore preserve the integrity of the memory device.
In two-level memories, since the distance, in terms of difference in output signal, between the threshold of a virgin cell and the threshold of a programmed cell is rather large, it is not necessary to have, as in the previous case, a reference cell with a threshold that lies between the threshold of a virgin cell and the threshold of a programmed cell; the same effect is achieved by using a virgin cell as reference and by unbalancing the two branches of the sense amplifier.
In a multilevel memory, each individual cell must instead contain a plurality of logic states: for example, in a four-level memory there are four logic states, "11", "10", "01" and "00", which are mutually separated by three reference levels (thresholds), designated respectively by Vr1, Vr2 and Vr3 in FIG. 1, which are used to discriminate the four logic states during reading.
FIG. 1 charts the threshold of a cell as a function of the number of cells.
The above-mentioned logic values that the cell can assume are mutually separated by the above-mentioned three references together with the supply voltage Vcc.
The condition in which the cell is erased in an optimum manner corresponds to the logic state "11", in which the threshold of the cell is below the threshold of the lowest reference Vr1.
The curve plotted in the chart of FIG. 1 shows the distribution of virgin cells.
For technological reasons, the choice of the references should not be made in advance, i.e., regardless of the device being considered, as instead occurs in a two-level memory.
Erasure of the matrix cells by means of ultraviolet rays in fact produces a statistical distribution of their thresholds (as shown in FIG. 1), and since the above-described testing method eliminates memory cells whose threshold is higher than the reference Vr1, an excessively low threshold Vr1 entails the risk of eliminating an excessive number of memory cells which are in the upper trailing part of the statistical distribution, so that the redundancy lines are not sufficient to replace all the eliminated memory cells.
Vice versa, an excessively high threshold Vr1 reduces the interval between said threshold and the supply voltage Vcc, in which the upper references of Vr2 and Vr3 must be defined.
Therefore, since not only the thresholds of virgin cells but also the thresholds of cells written with a given logic value have a certain statistical distribution with respect to their average value, a reduction of the intervals above the minimum threshold Vr1 that define the logic level entails an increase in the probability of an error when reading the cell.
The narrower the intervals between the logic write levels, the higher the likelihood that said logic levels will not be discriminated with sufficient precision.